Velum Labs

What do they actually do

Velum Labs provides a privacy-first AI platform that lets organizations run machine learning on encrypted data using fully homomorphic encryption. Customers encrypt data locally with Velum’s SDK, send ciphertext to Velum’s APIs for indexing/inference/training, and decrypt results on their side so plaintext never leaves their environment. The service is in private/early beta with access by request and enterprise-style legal terms (DPA) available (platform overview, early access, DPA).

They also describe a content-level “firewall” that enforces access policies across both human users and AI systems; this appears as a forward-looking complement to the encrypted-compute platform, with integration details still limited in public materials (YC listing, Velum site).

Who are their target customer(s)

• Hospitals, health systems, and medical labs: Need to run analytics and models on protected health information without exposing plaintext or violating HIPAA-like obligations; require auditability and strong data-processing agreements (inference page, DPA).
• Banks, insurers, and trading firms: Must analyze sensitive customer and risk data while meeting strict regulatory and audit requirements; cannot send raw data to third-party services and need encrypted ML workflows (inference page, DPA).
• Government agencies and public-sector bodies: Handle classified/PII data and require residency controls, auditability, and options beyond public cloud; want ML capabilities without exposing plaintext to vendors (inference page, YC listing).
• Enterprise security, privacy, and compliance teams: Struggle to enforce who or what (humans and AIs) can see sensitive content across systems; need technical controls and audit trails to prevent leaks and satisfy regulators (YC listing, DPA).
• SaaS product teams building AI on customer data: Want to offer AI features without requiring customers to upload plaintext to external LLMs; risk losing deals without strong privacy guarantees and simple SDKs/APIs (early access, Velum site).

How would they acquire their first 10, 50, and 100 customers

• First 10: Run high‑touch pilots with a small set of hospitals, banks, and government teams sourced from early‑access signups and founder/YC networks; embed engineers to integrate the SDK, sign the DPA up front, and do limited custom work to prove the encrypt→API→decrypt flow in production‑like settings (early access, DPA).
• First 50: Package a 30–90 day standardized pilot with fixed deliverables and compliance artifacts; scale via referrals plus partnerships with security/privacy consultancies and vertical software vendors, producing short case studies and repeatable procurement playbooks (platform overview, DPA).
• First 100: Productize onboarding and SDKs for self‑serve, publish connectors/policy rules for common enterprise systems, and sign reseller/channel agreements with cloud/SIs while investing in audit‑ready documentation to clear enterprise procurement at scale (Velum site, YC listing).

What is the rough total addressable market

Top-down context:

Specialist homomorphic‑encryption markets are currently in the low hundreds of millions, while broader privacy‑enhancing technologies (PETs) are several billion today and projected toward ~USD 12B by 2030; regulated AI spend in healthcare/finance is already tens of billions, providing long‑term context if encrypted compute becomes standard (GMI on FHE, Grand View on PETs, AI in healthcare, AI in finance).

Bottom-up calculation:

Illustratively, if 500–1,500 regulated enterprises adopt encrypted‑ML pilots at USD 150k–300k ACV for a small number of workloads, that implies a near‑term SAM of roughly USD 75M–450M; expanding usage across more workloads and sectors could grow this into the low billions as PET adoption broadens.

Assumptions:

• Initial ACVs in the USD 150k–300k range for enterprise pilots and early production workloads.
• 500–1,500 near‑term target buyers across healthcare, finance, and public sector with budget for PETs.
• Adoption ramps gradually as performance, integrations, and certifications mature.

Who are some of their notable competitors

• Zama: Open‑source FHE tooling (Concrete ML) to convert/run ML models on encrypted data; strong with developers/researchers and overlaps with Velum’s encrypted inference/training focus (docs, GitHub).
• Duality Technologies: Commercial data‑collaboration and PETs platform using FHE and related techniques for secure analytics/ML in regulated industries; competes on enterprise features and compliance (overview, ML).
• Enveil: Sells “data‑in‑use” protections and secure AI/search so organizations can process sensitive data without exposing plaintext; targets similar bank/healthcare/government use cases (site, data in use).
• Inpher: “Secret computing” (MPC‑style) products for cross‑organization analytics and ML without sharing raw data; alternative technical approach that competes for similar enterprise budgets (products, technology).
• Microsoft Azure Confidential Computing: Hardware‑based confidential computing on a major cloud; production‑ready option some enterprises may choose instead of FHE for protecting data while in use (overview).